In this paper, we address the challenging task of simultaneously optimizing (i) the weights of a neural network, (ii) the number of neurons for each hidden layer, and (iii) the subset of active input features (i.e., feature selection). While these problems are traditionally dealt with separately, we propose an efficient regularized formulation enabling their simultaneous parallel execution, using standard optimization routines. Specifically, we extend the group Lasso penalty, originally proposed in the linear regression literature, to impose group-level sparsity on the network's connections, where each group is defined as the set of outgoing weights from a unit. Depending on the specific case, the weights can be related to an input variable, to a hidden neuron, or to a bias unit, thus performing simultaneously all the aforementioned tasks in order to obtain a compact network. We carry out an extensive experimental evaluation, in comparison with classical weight decay and Lasso penalties, both on a toy dataset for handwritten digit recognition, and multiple realistic mid-scale classification benchmarks. Comparative results demonstrate the potential of our proposed sparse group Lasso penalty in producing extremely compact networks, with a significantly lower number of input features, with a classification accuracy which is equal or only slightly inferior to standard regularization terms

Neural networks, as powerful tools for data mining and knowledge engineering, can learn from data to build feature-based classifiers and nonlinear predictive models. Training neural networks involves the optimization of nonconvex objective functions, and usually, the learning process is costly and infeasible for applications associated with data streams. A possible, albeit counterintuitive, alternative is to randomly assign a subset of the networks' weights so that the resulting optimization task can be formulated as a linear least-squares problem. This methodology can be applied to both feedforward and recurrent networks, and similar techniques can be used to approximate kernel functions. Many experimental results indicate that such randomized models can reach sound performance compared to fully adaptable ones, with a number of favorable benefits, including (1) simplicity of implementation, (2) faster learning with less intervention from human beings, and (3) possibility of leveraging overall linear regression and classification algorithms (e.g., ℓ 1 norm minimization for obtaining sparse formulations). This class of neural networks attractive and valuable to the data mining community, particularly for handling large scale data mining in real-time. However, the literature in the field is extremely vast and fragmented, with many results being reintroduced multiple times under different names. This overview aims to provide a self-contained, uniform introduction to the different ways in which randomization can be applied to the design of neural networks and kernel functions. A clear exposition of the basic framework underlying all these approaches helps to clarify innovative lines of research, open problems, and most importantly, foster the exchanges of well-known results throughout different communities.

Neural networks are generally built by interleaving (adaptable) linear layers with (fixed) nonlinear activation functions. To increase their flexibility, several authors have proposed methods for adapting the activation functions themselves, endowing them with varying degrees of flexibility. None of these approaches, however, have gained wide acceptance in practice, and research in this topic remains open. In this paper, we introduce a novel family of flexible activation functions that are based on an inexpensive kernel expansion at every neuron. Leveraging over several properties of kernel-based models, we propose multiple variations for designing and initializing these kernel activation functions (KAFs), including a multidimensional scheme allowing to nonlinearly combine information from different paths in the network. The resulting KAFs can approximate any mapping defined over a subset of the real line, either convex or nonconvex. Furthermore, they are smooth over their entire domain, linear in their parameters, and they can be regularized using any known scheme, including the use of 1 penalties to enforce sparseness. To the best of our knowledge, no other known model satisfies all these properties simultaneously. In addition, we provide a relatively complete overview on al- * ternative techniques for adapting the activation functions, which is currently lacking in the literature. A large set of experiments validates our proposal.

No abstract

No abstract